Graphcore’s Bow-2: A Serious Challenger in AI Accelerators

Graphcore just launched the Bow-2, their new AI accelerator, and it’s got some genuinely disruptive hardware ideas. The Bow-2’s big trick is stacking compute and memory wafers using a wafer-on-wafer (WoW) process—think TSMC’s SoIC, but with Graphcore’s IPU twist. The result? Memory bandwidth that rivals Blackwell, and compute density that’s honestly wild.

WoW + HBM4: Not Just Buzzwords

Why should engineers care? First, Bow-2 integrates HBM4 directly onto the logic wafer, making data movement a non-issue for large models. Memory bottlenecks are the #1 headache in training and inference at scale. Second, the WoW approach means lower latency and higher power efficiency, which is critical when you’re deploying clusters in the real world—not just racking up FLOPs in a datacenter showcase.

A Real Threat to the Status Quo?

For the first time, NVIDIA and AMD have credible competition outside the US-Asia duopoly. Graphcore’s new SDK supports PyTorch and JAX out-of-the-box, so you’re not stuck rewriting model code. In my opinion, this is the most exciting thing Graphcore has done since the original IPU. Anyone building custom inference infrastructure—or worried about vendor lock-in—should pay close attention. This is how innovation happens: not by chasing bigger numbers, but by rethinking the stack.